Shipwrecks Detection Based on Deep Generation Network and Transfer Learning with Small Amount of Sonar Images

Abstract

The application of deep learning sonar target detection is severely limited due to the small amount of sonar images, especially for submarine shipwreck. Aiming to overcome the over-fit of training problem and improve accuracy of detection, we proposed a method which combine deep generation networks and transfer learning for sonar shipwrecks detection. Specifically, in deep generation network, we used similarity measurement to improved optimization, which generate high quality fake image and laid the further foundation of data. Then, in transfer learning detection, we used multi-layer adaptation and multi-core MMD to fine-tune and frozen pre-trained model, prevent the problem of over-fit and improve the generalization and stability of the system. And we combined the methods of regional suggestion and regression for target detection to guarantee precision of detection. Finally, the contrast experiment of sonar shipwrecks is carried out the effectiveness of the proposed method.